- EMNLP 2022
The scientific method is humanity’s most powerful and successful process for advancing our knowledge and solving complex problems. The process always starts with a question, before any research can begin to then form a hypothesis. In order to formulate a hypothesis or improve upon an earlier take, each new piece of downstream information should be fed back into the earlier stages of the process. In materials discovery, this approach is expensive and typically takes years to come up with a new, viable material. Nearly every step is peer reviewed to ensure the consistency and reliability of the results. Current processes frequently lack tools of traceability to ensure that decisions can be traced back to all relevant data that informed each outcome.
At IBM Research, we’re working to provide technological solutions that assist subject matter experts (SMEs) with their scientific workflows by (i) enabling more effective and scalable knowledge representation, (ii) rendering AI technologies to participate equally in the scientific process, and (iii) facilitating intelligent and effective lab automation.
Knowledge is the biggest resource consumed and created during the discovery process. We are looking into methods and enabling technologies to make knowledge representation and curation for the discovery of complex materials more effective and scalable. We incorporate a variety of knowledge sources, including scientific literature, patents, experiments, and the inherent expertise of experts, throughout the intake process. As a result, we are starting to develop novel representations of complex materials, such as polymers, to contribute to the scientific community.
Due to AI's rapid evolution, the concept of a peer in the scientific community must inevitably be expanded to take into account various AI agents. We are researching mechanisms and developing tools to enable AI agents to participate equally in the scientific process and be subject to peer review. We concentrate on the patterns and workflows that make use of RoboRXN, AI-augmented simulations, and generative models—all key components of accelerated discovery. Our approach to creating expert-AI interactions in science focuses on how AI can support SMEs and how peer review works between SMEs and AI. For example, any proposed candidate by AI must undergo a thorough evaluation before moving on to the experimental stage. During this evaluation, the SME needs a method for requesting proof points about the viability and value of the proposed candidates from the AI agent. We are creating systems to progressively support the assessment stage while also enhancing the explainability and traceability of ideas produced by AI.
Given the urgency, we are initially focused on addressing the environmental and human health impact of PFAS ‘forever chemicals’ by finding sustainable replacements and improved capture materials. Although the technologies we are developing are valuable and applicable for any form of scientific discovery.
- NeurIPS 2022
- ACS Fall 2022
- ACS Fall 2022
- ICML 2022
- Development of polymer information systems to support accelerated materials development via AI-enabled automated experimentation
- ACS Spring 2022
- Development and use of a software platform applied towards the automated synthesis of carbonate monomers and machine learning applications
- ACS Spring 2022
- Large datasets for machine learning and reactor response curves for the ring opening polymerization of lactide
- ACS Fall 2021
Sustainable replacements for PFAS
- Materials Discovery
- Accelerated Discovery